# Claude Opus 5.5: Separating the 40% Savings Claim from the 20% Token Price Cut
**Plutonous** | September 26, 2026 | 6 min read
> Opus 5.5 lowers standard token prices by 20%, while Anthropic estimates 40% lower costs for typical tasks. The difference matters for budgets and API migrations.
Tags: Claude Opus 5.5, Anthropic, AI Pricing, Claude API, Prompt Caching, AI Agents, Model Migration, Developer Tools
---
**TL;DR: Anthropic launched Claude Opus 5.5 on September 22 at $4 per million input tokens and $20 per million output tokens, each 20% below Opus 5. Cache reads fell 60%, from $0.50 to $0.20 per million.[1] Anthropic's roughly 40% lower typical-workload cost is a vendor estimate that also relies on token efficiency and workload mix; it is not a universal invoice discount. Migration changes to thinking, tools and response handling deserve a staged test.[4]**
The real story isn't the price tag alone. Anthropic has made its premium model cheaper per token while shifting more of the economic case onto how much work an agent completes per token. That is attractive for long coding sessions. It is also precisely why a procurement team should demand task-level measurements before writing 40% into a forecast.
The September 22 release is now available through the Claude API and major cloud platforms; the API model ID is `claude-opus-5-5`.[1][2] This September 26 analysis concerns a four-day-old release, not a new launch today. Anthropic's own comparison says Opus 5.5 performs at Fable 5.1's level on most work. That is a vendor characterization, not an independently normalized benchmark result.[1]
### Why This Matters Now
The published list price and the estimated cost per completed task answer different questions. A migration can lower the bill, raise it on a particular workload, or fail at request validation. Test both economics and compatibility before routing production traffic.
*Cover image: generated editorial artwork of an antique calculator, blank ledger and trays of counters. It is a metaphor for accounting and workflow, not Anthropic hardware, a measured cost chart or benchmark evidence.*
## The Price Sheet: 20% Is the Guaranteed Rate Change
Anthropic lists Opus 5 at $5 per million input tokens and $25 per million output tokens. Opus 5.5 lists $4 and $20. A five-minute cache write moves from $6.25 to $5 per million; a cache read moves from $0.50 to $0.20.[1][3] The input, output and five-minute write rates therefore fall exactly 20%; cache reads fall 60%. These are Claude Platform token rates, before platform-specific arrangements, credits, taxes or other features.
| Billed category, per million tokens | Opus 5 | Opus 5.5 | Rate change |
| --- | ---: | ---: | ---: |
| Ordinary input | $5.00 | $4.00 | -20% |
| Output | $25.00 | $20.00 | -20% |
| Five-minute cache write | $6.25 | $5.00 | -20% |
| Cache read | $0.50 | $0.20 | -60% |
Here's the genius in the pricing: cache reads become much cheaper just as agent workflows repeatedly revisit the same large context. But a lower read price only helps when the request actually hits the cache. Anthropic documents prefix and lifetime requirements; changing a cached prefix can turn a cheap read into a write or ordinary input charge.[5] The operational metric is the billed cache-read count, not whether caching is switched on. Our [prompt-caching cost guide](/news/claude-prompt-caching-cost-break-even) explains the break-even arithmetic for repeated context.
## The 40% Claim: Task Efficiency Is Doing Extra Work
Anthropic says Opus 5.5 costs about 40% less than Opus 5 on typical workloads billed by token. Its explanation combines the lower rates with fewer tokens per task; that is based on Anthropic's tests and cannot be assumed for every customer's prompts, effort settings or cache behavior.[1] At unchanged token counts with no cache, the arithmetic is only 20% savings.
Consider a **worked editorial example, not measured production usage**. A completed task consumes 1 million ordinary input tokens and 100,000 output tokens on Opus 5. At published rates, that is $5 + $2.50 = **$7.50**. If Opus 5.5 consumes exactly the same tokens, it costs $4 + $2 = **$6.00**, a 20% reduction. For that task to cost $4.50, or 40% below $7.50, the new model would need a different token mix: for example, 750,000 ordinary input and 75,000 output tokens, costing $3 + $1.50. That example assumes both counts fall 25%, with no cache, batch or fast-mode charges. It proves the arithmetic, not that the model will deliver that reduction.
What's often overlooked is the denominator. Agents can retry, call tools, spawn subagents and consume thinking tokens. Thinking tokens are billed as output even when their text is omitted from the response.[4] Measure cost per **accepted completed task**, including failed attempts and tool calls, alongside quality and elapsed time. A lower invoice per token can coexist with a higher invoice per accepted result.
## Migration: Four Compatibility Checks Before the Model Swap
The model ID change is one line; the behavioral contract is wider. Anthropic's migration guide says Opus 5.5 always uses adaptive thinking. Requests that disable thinking or specify a manual thinking budget return HTTP 400. Explicit `effort` is useful because the default moves from Opus 5's `high` to Opus 5.5's `medium`; an unchanged request would otherwise compare different reasoning settings.[4][7]
Second, forced `tool_choice` values `any` and named `tool` are rejected. Use `auto` and validate required output with strict tools or structured outputs. Third, read response blocks by `type`, not by assuming `content[0]` is text. Preserve signed thinking blocks unmodified through tool loops and test any router that changes models mid-conversation. Fourth, Claude API and Google Cloud integrations using `computer_20251124` must adopt `computer_toolset_20260801`; Anthropic says the older tool remains compatible on Amazon Bedrock.[4]
A practical rollout is small and measurable: inventory payloads and response parsers, run a representative evaluation set at explicit effort levels, compare bills per accepted task, then canary traffic with a rollback route. Include a long cached agent session, a forced-tool workflow, a streamed response and a computer-use flow if those exist in production. Anthropic recommends development testing before production switching.[4]
## Speed and Capacity: Price the Mode Actually Used
Fast mode is a separate choice and a research preview on the Claude API that requires access. Anthropic advertises up to 2.5 times higher output tokens per second, but charges $8 per million input and $40 per million output tokens on Opus 5.5, twice the standard Opus 5.5 rates.[1][8] The vendor-reported speed figure concerns output generation, not time to first token, and is no latency promise for a particular prompt or concurrency level. Likewise, the standard 1-million-token context window and 128,000-token maximum output describe capacity, not a recommended request size or a fixed cost.[2]
The uncomfortable truth is that a 40% headline can produce poor budgeting if teams copy it straight into financial plans. Separate standard and fast-mode traffic, cache writes and reads, and retries. Anthropic's pricing documentation also lists a 50% Batch API discount on input and output, which is a separate processing choice rather than part of the 40% estimate.[2][6]
### The Key Insight
Use 20% as the published ordinary input/output rate reduction. Treat roughly 40% as Anthropic's typical-workload estimate and replace it with your own accepted-task measurement before changing budgets or default routing.
Let's be clear: Opus 5.5 gives buyers a real lower unit price. The larger business advantage must be earned in a specific workflow, with compatible requests and verified outcomes. That is where the migration decision belongs.
## Sources
1. [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5)
2. [Claude Opus 5.5 overview](https://platform.claude.com/docs/en/models/opus-5-5/overview)
3. [Claude Opus 5 overview](https://platform.claude.com/docs/en/models/opus-5/overview)
4. [Migrating to Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/migration-guide)
5. [Prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching)
6. [Pricing](https://platform.claude.com/docs/en/about-claude/pricing)
7. [Effort](https://platform.claude.com/docs/en/build-with-claude/effort)
8. [Fast mode](https://platform.claude.com/docs/en/build-with-claude/fast-mode)
*Last updated: September 26, 2026*
---
*Source: [LLM Rumors](https://www.llmrumors.com/news/claude-opus-55-pricing-migration)*